Plan d'implantation de Télécom ParisTech, de l'Institut Télécom et ...
Modeling the Energy Consumption of Programs: Thermal ... · Approach for Platform Energy...
Transcript of Modeling the Energy Consumption of Programs: Thermal ... · Approach for Platform Energy...
Institut Mines-Télécom
Modeling the Energy Consumption of
Programs: Thermal Aspects and
Energy/Frequency Convexity Rule
Karel De Vogeleer, Hypervirtu
Kameswar Rao Vaddina, TelecomParisTech, U. Paris-Saclay
Florian Brandner, TelecomParisTech, U. Paris-Saclay
Pierre Jouvelot, MinesParisTEch, PSL U.
Gérard Memmi, TelecomParisTech, U. Paris-Saclay
WCSP'2017 1
Institut Mines-Télécom
Why focussing on energy saving for
mobile computing?
It is not only about the magnitude of saved energy: • A smartphone CPU consumes between 60 to 500mW
• There were about 7x109 smartphones sold in the last 5 years, there will be 50x109 ‘smart objects’ in 2022
• A worldwide saving of 30% would roughly mean about 280 MW for the smartphones, about 3 GW for the smart objects
• This would ‘only’ save between one tidal and one nuclear power station worldwide
Saving energy at the software level also is about a natural-resource-free energy saving
Focussing on mobile systems (e.g. a baystation on a drone): they are ‘energy-critical’ : it is about being constantly looking for providing more autonomy with an unchanged QoS, with the same battery
WCSP'2017
2
Institut Mines-Télécom
The trend is moving towards providing more
computational power
WCSP'2017 3
Institut Mines-Télécom
Thermal Behavior:
Power-temperature rule
Passive Cooling rule
WCSP'2017 4
Institut Mines-Télécom
Temperature impacts energy consumption
Temperature Power
WCSP'2017
2/
1 0
T aP a e a
An increase of 10% of temperature generates
an increase of 5% in power consumption
5
Institut Mines-Télécom
Small size mobiles have no fan
Passive Cooling Rule
WCSP'2017
Approximations do exist for small areas
6
Institut Mines-Télécom
Contributions on thermal behavior
Necessary for reproducible measurements and for accurate energy consumption models
Power–temperature relationship
Approximations for practical uses, particularly for online usages (embedded systems, radio mobiles,…)
WCSP'2017 7
Institut Mines-Télécom
EFCR: the energy – frequency
convexity rule
WCSP'2017 8
Institut Mines-Télécom
Fragmenting energy consumption per
system module
WCSP'2017
sufficiently constant over time
9
Institut Mines-Télécom
Power and time model
WCSP'2017 10
Institut Mines-Télécom
V can be approached by a linear function of
the frequency
WCSP'2017
V = m1 f + m2
11
Institut Mines-Télécom
Optimal frequency and Convexity
(EFCR)
WCSP'2017 12
Institut Mines-Télécom
State of the art
Convexity was already observable, however no
analytical studies were performed
Fan, X., Ellis, C. S., and Lebeck, A. R.
The synergy between power-aware memory systems and processor voltage scaling. In PACS’04
Le Sueur, E., and Heiser, G.
Dynamic voltage and frequency scaling: the laws of diminishing returns. In PACS’10
WCSP'2017 13
Institut Mines-Télécom
Convexity shown on Intel Core 2 board
R. Efraim, R. Ginosar, C. Weiser, &A. Mendelson: “Energy Aware Race to Halt A down to EARtH
Approach for Platform Energy Management”, IEEE Computer Architecture Letter, 2012.
WCSP'2017 14
Institut Mines-Télécom
Two Testbeeds
WCSP'2017
Host running Labview
NI Monitoring Hardware
TI AM572x Board under test
More expensive
More accurate
Probes location anywhere on the
board
Various SW analytics
“Experimental Energy Profiling of Energy-critical Embedded Applications”
K. Vaddina, F. Brandner, P. Jouvelot, and G. Memmi
IEEE SoftCom’17, Split, Croatia, 2017.
Cheaper
Easier to set up, easier to use
Probes at the battery leve
Home made analytics
Samsung Galaxy SII under test
Monsoon Monitoring Hardware
Host running Analytics
15
Institut Mines-Télécom
Experimental validation with the
Samsung smartphone
In color: measurements
In doted lines: theoretical EFCR calculation
When N increases, fopt stays stable
WCSP'2017 16
Institut Mines-Télécom
Experimentation with TI AM572x board
WCSP'2017 17
Institut Mines-Télécom
fopt sensitivity
Energy consumption of three different programs running on TI
AM572x platform showing different profiles with different fopt.
WCSP'2017
“Parameter Sensitivity Analysis of the Energy/Frequency Convexity Rule for Application Processors”
K. De Vogeleer, G. Memmi, and P. Jouvelot
J. of Sustainable Computing, Informatics and Systems, Elsevier B.V., September 2017.
18
Institut Mines-Télécom
Towards program energy
profiling
Institut Mines-Télécom
Energy profiles also can detect anomalies
WCSP'2017 20
Institut Mines-Télécom
0.8 1.0 1.2 1.4 1.6
0.0
40
.05
0.0
60.0
70
.08
0.0
90
.10
0.1
1
dijkstra
frequency (GHz)
en
erg
y (J
)
−O0
−O1
−O2
−O3
Optimizing for performance also optimizes
for energy cpu
WCSP'2017 21
Institut Mines-Télécom
Drawing the best from the optimizer and DVFS
• Created by tuning clk frequency and performing standard program transformation
WCSP'2017 22
Institut Mines-Télécom
Conclusion
WCSP'2017 23
Institut Mines-Télécom
First measurements and results are setting expectations in the 10-40% saving
range by:
Exploiting energy-frequency convexity
Integrating temperature impact in our models
Energy-Oriented Environment
Wider array of experimentation
• Using a wider and better controled temperature range
• Setting a richer and more complete benchmark
More research on energy program profiling
• Handling various architectures (e.g. cache)
• Understanding how where, and when to play with clock frequency changes (including overhaed data)
• Temperature online monitoring
WCSP'2017 24
Institut Mines-Télécom
Thank you
Institut Mines-Télécom
Bibliography
“Experimental Energy Profiling of Energy-critical Embedded Applications” K. Vaddina, F. Brandner, P. Jouvelot, and G. Memmi IEEE SoftCom’17, Split, Croatia, 2017.
“Parameter Sensitivity Analysis of the Energy/Frequency Convexity Rule for Application Processors” K. De Vogeleer, G. Memmi, and P. Jouvelot, J. of Sustainable Computing, Informatics and Systems, Elsevier, September 2017.
“Modélisation de la consommation énergétique des programmes : aspects thermiques et loi de convexité énergie-fréquence” K. De Vogeleer, P. Jouvelot, and G. Memmi, ICSSEA’16 then Génie Logiciel 117 pp 47-59, June 2016.
“Modeling Temperature Bias of the Power Consumption of Nanometer-Scale CPUs in Application Processors.” K. De Vogeleer, G. Memmi, P. Jouvelot, and F. Coelho International Conference on Embeded Computer Systems: Architectures, Modeling, and Simulation, SAMOS XIV, July 2014.
“The Energy/Frequency Convexity Rule: Modeling and Experimental Validation on Mobile Devices” K. De Vogeleer, G. Memmi, P. Jouvelot, and F. Coelho 10th International Conference on Parallel Processing and Applied Mathematics, PPAM 2013, PEAC Workshop on "Power and Energy Aspects of Computation", Warsaw, Poland, pp 793-803, September 2013.
WCSP'2017 26